Remote Sensing Image Scene Classification Based On Deep Transfer Learning

With the advancement of imaging technology,remote sensing image obtained by sensors have a higher spatial resolution than before.At present,remote sensing images have been used in many research domains,and scene classification is one of the hot topics.However,there are some constraints in the scene classification task that limit the classification performance,including the limited training samples,the limitation of lowlevel feature-based classification methods,and the intra-class diversity among images.The above scene classification is carried out in a remote sensing dataset with the fullysupervised manner,that is,the training set and test set are from the same dataset.When performing the scene classification on a new dataset without label information,that is,the training set and test set are from different dataset,the cross-domain scene classification is required.To solve the above issues,we study the feasibility of deep transfer learning in remote sensing image scene classification task in this paper.When the training and test samples come from the same dataset,we use the model pre-trained on Image Net,and combine it with label augmentation and intra-class constraint to improve the classification performance.When the training and test samples come from different datasets,we use the domain adaptation to reduce the discrepancy of different domains.To sum up,the contents of this paper are in three aspects:Firstly,in the case that the training and test samples come from the same dataset,since the number of training samples in remote sensing datasets is generally small,we use pre-trained model to reduce the demand of network on training samples.In order to fully use the training samples,the label augmentation is proposed to fully use the training samples,and the Kullback–Leibler divergence(KL)is used to constrain the output distribution of two remote sensing images with the same scene category to reduce the intra-class diversity.In addition,we combine label augmentation and intra-class constraint to further improve the classification performance.Then,for the situation that the training samples and test samples are from different dataset,we divide it into single-source domain adaptation and multi-source domain adaptation.For the single-source domain adaptation,the training samples are from the same dataset.We use the local maximum mean discrepancy to divide similar samples into a sub-domain,and then align the features in each sub-domain.Finally,for the multi-source domain adaptation,the training samples are from different dataset.We mainly study the situation that the training samples come from two dataset,that is,there are two source domains.To complete the multi-source domain adaptation,we align the feature distribution and the network output respectively.Compared with the single source domain,the classification performance can be improved by reasonably using samples from multiple source domains.